Timestral
What it is
Timestral is a family of four time series foundation models designed for probabilistic forecasting, ranging from 5M to 209M parameters. It utilizes Decoupled Horizon Forecasting to separate prediction tokens from context, enabling efficient training on longer horizons, and employs a two-stage pipeline involving synthetic pretraining followed by mixed real-synthetic post-training.
The problem it solves
Existing time series models struggle with long-horizon forecasting due to context masking issues and lack native support for known-future covariates within causal architectures.
What you could use it for
- Probabilistic time series forecasting
- Benchmarking against state-of-the-art models on GIFT-Eval and TIME
- Forecasting with known-future covariates
Why it matters
Establishes a new state of the art in probabilistic time series forecasting, with the 123M and 209M parameter models ranking in the top two positions on major benchmarks.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.